Predicting POC export from plankton diversity

Calipsocean meeting

Thelma Panaïotis

05/03/2024

Aim

Plankton & carbon

  • Role of plankton in C cycle

  • C export mostly estimated through modeling

  • Limited representation of zooplankton diversity

Can we use observations and machine learning to improve the representation of zooplankton diversity in ESM?

Questions

env env plankton* plankton* env->plankton* carbon export** carbon export** env->carbon export** plankton*->carbon export**

  • Can POC be predicted from env?
  • Can POC be predicted from plankton diversity?
  • Can plankton diversity be predicted from env?

*plankton = zooplankton **Here, we will use POC 1000 m as a proxy for carbon export.

Data

POC

POC export at 1000 m

Environment

As in Wang et al., 2023.

Yearly climatologies from GLODAPv2.

  • temperature

  • silicate

  • phosphate

  • oxygen

  • NPP

  • alkalinity

  • DIC

  • DOC

Plankton

UVP5 dataset: 2876 profiles

Plankton - Taxonomy

  • taxonomic richness

  • taxonomic diversity

  • taxonomic evenness

Plankton - Morphology

Functional diversity metrics (Magneville et al. 2022)

→ morphological diversity metrics (Beck et al., 2023)

  • morphological richness

  • morphological divergence

  • morphological evenness

Plankton - Trophic Status (not yet)

The problem(s)

  • low taxonomic resolution of the UVP5 dataset
  • trophic status of a chaetognath
  • trophic status of a copepod?

Plankton - Trophic Status (not yet)

The ideas

  • by size?
  • other dataset? ↑ taxonomic resolution but ↓ coverage.
  • plankton NASS to infer trophic structure?

Final dataset

Placeholder

Placeholder

439 data points

Final dataset – Splits

Training* VS test set, stratified by POC.

TODOs

  • Account for spatial autocorrelation (spatial CV)

  • Get more robust estimates of R² (nested CV)

ML Model

(Multivariate) boosted trees

Response variable:

  • uni- or multivariate

  • ~normally distributed → log(POC)

Flexibility for predictors, handles interactions.

Complex & non-linear relationships.

Easy interpretation & implementation.

Results

POC from env

POC ~ temperature + silicate + phosphate + oxygen + NPP + alkalinity + DIC

R² = 91.0%

Good prediction!

POC from plankton

POC ~ all plankton metrics

R² = 57.1%

OK prediction!

Best predictors:

  • ta. richness ×2

  • mo. richness

POC from plankton best predictors

POC ~ ta_ric_3 + ta_mast + mo_ric

R² = 38.8%

OK prediction!

POC from plankton best predictors

POC ~ ta_ric_3 + ta_mast + mo_ric

POC response to plankton descriptors.

TODO: Merge both descriptors of taxonomic richness into one.

plankton best predictors from env

ta_ric_3 + ta_mast + mo_ric ~ temperature + silicate + phosphate + oxygen + NPP + alkalinity + DIC

Mult. R² = 37.8%

OK prediction!

Conclusion

Answers

  • Can POC be predicted from env? Yes
  • Can POC be predicted from plankton diversity? Yes
  • Can plankton diversity be predicted from env? Yes

Take-home messages

  • We can predict 38.8% of POC variance using only 3 descriptors of the zooplankton community.
  • We know how POC reponds to variations in these plankton descriptors.
  • These plankton descriptors can be predicted from the environment.